2021 IEEE International Conference on Robotics and Automation (ICRA 2021)

May 31 - June 4, 2021, Xi'an, China

High-Speed Robot Navigation

using Predicted Occupancy Maps
Kapil D. Katyal1,2 , Adam Polevoy1 , Joseph Moore1 , Craig Knuth1 , Katie M. Popek1
2021 IEEE International Conference on Robotics and Automation (ICRA) | 978-1-7281-9077-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICRA48506.2021.9561034

Abstract— Safe and high-speed navigation is a key enabling

capability for real world deployment of robotic systems. A sig-
nificant limitation of existing approaches is the computational
bottleneck associated with explicit mapping and the limited
field of view (FOV) of existing sensor technologies. In this
paper, we study algorithmic approaches that allow the robot
to predict spaces extending beyond the sensor horizon for
robust planning at high speeds. We accomplish this using a
generative neural network trained from real-world data without
requiring human annotated labels. Further, we extend our
existing control algorithms to support leveraging the predicted
spaces to improve collision-free planning and navigation at high
speeds. Our experiments are conducted on a physical robot
based on the MIT race car using an RGBD sensor where were
able to demonstrate improved performance at 4 m/s compared
to a controller not operating on predicted regions of the map.
I. I NTRODUCTION
A key objective of mobile robotic systems is to execute
safe, reliable motion while avoiding obstacles in the shortest
amount of time possible. While mobile robots have demon-
strated considerable success in recent years, they still fail to
maneuver in environments at the speed and agility of hu-
mans. Traditional navigation algorithms require explicit per-
ception, mapping, localization and control for collision free
motion. Often, high-speed navigation using these traditional
approaches is severely limited by the sensor’s field of view
(FOV) as well as the computational requirements needed for
explicit mapping. This requires a robot to frequently reduce
its speed to rescan the environment, construct a map and
replan a new trajectory.
This is in contrast to human navigation where cognitive
psychologists have hypothesized that humans (i) actively
differentiate between occupied and free spaces based on ob-
servations, (ii) make predictions of occupied spaces beyond
line of sight and (iii) use these predictions for navigation to
help improve robustness and agility [1], [2], [3].
In this paper, we lay the foundation for developing al-
gorithms that provide these capabilities to robotic systems.
Our intuition is that as humans navigate, they leverage spatial
cues within the environment to generate predictions of future
spaces and use that as part of the planning process. Our
objective is to mimic this predictive capability in robotic
systems (see Fig. 1). Our approach similarly learns to predict
spaces beyond the line of sight and further uses the predicted
areas as part of the robot controller for planning.
1 Johns Hopkins University Applied Physics Lab, Lau-
rel, MD, USA. Kapil.Katyal@jhuapl.edu,
Adam.Polevoy@jhuapl.edu
2 Dept. of Comp. Sci., Johns Hopkins University, Baltimore, MD, USA.
Fig. 1: A motivating example of our approach. We leverage
occupancy map prediction to allow the robot to plan more
robust trajectories at higher speeds compared to those limited
by the sensor horizon.
Our specific contributions include:
• Novel perception algorithms that predict future occu-
pancy maps using generative neural networks.
• A controller that leverages the predicted occupancy map
during planning.
• Real-world hardware experiments using a robotic car
to demonstrate higher speed navigation with improved
reliability compared to a controller not operating on
predicted regions of the map.
II. R ELATED W ORK
Previous work has presented different strategies to predict
unknown parts of the map, e.g. [4]. A variety of methods
incorporate map predictions to speed exploration of an envi-
ronment [5], [6], [7], [8], [9]. Predominantly, these methods
are applied to exploration and therefore do not stress the
capability of the map prediction module with respect to the
control pipeline. In particular, this paper extends our prior
work, [7], by using a balanced classification loss instead
of a regression loss in addition to data augmentation and
noise suppression techniques required to generate accurate
predictions using real data.
An alternative to map prediction is presented in [10]; this
method predicts the next best viewpoint for exploration by
utilizing experience from previously mapped environments.
Our method shows that map prediction effectively increases
the amount of information available from the map and
improves point-to-point high speed navigation.

978-1-7281-9077-8/21/$31.00 ©2021 IEEE 5476

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
 Fig. 2: This diagram describes the overall perception and control pipeline. The perception module receives on-board sensor
data from the car and produces a predicted occupancy map using a U-Net style generative neural network. The control
algorithm receives the robot state, predicted occupancy map and goal point and generates collision-free trajectories.

Adjacent to map prediction is the problem of planning
the shortest path to a goal in an unknown environment.
These methods perform some level of inference over the en-
vironment to inform motion planning. For example, in [11],
they learn how to plan waypoints to a goal using a partial
map of the environment. Unlike our method, this approach
does not consider robot dynamics. Both [12] and [13] use
prior experience to reduce the size of viable environment
hypotheses. Specifically, [12] learns the probability of colli-
sion of motion primitives whereas [13] utilizes experience-
based map predictions in a belief space planner. Similar to
our approach, Elfhafsi et al. [14] incorporate map prediction
with global path planning that respects the system dynamics,
but all of their verification was done in simulation.
Existing work has also considered applying reinforcement
learning (RL) in tandem with occupancy maps or depth
information to navigate an unknown environment. In [15],
the method uses RGB images and predicts the probability and
variance of collision while navigating towards a goal, but is
incapable of global planning. In [16], the approach models
the world as a POMDP and take an end-to-end approach
to navigate to a goal. Additionally the method presented
in [17] uses a partial map as inputs to a deep RL policy
to navigate in unknown environments. While these methods
have shown success in application, the data requirements
of deep RL policies are large and often impractical in real
world scenarios. In addition, our method offers insight in the
process by intermediately predicting a map.
A key challenge in high-speed navigation in unknown
environments is balancing the speed of navigation against the
information gained from its sensors. Our method alleviates
this tension by learning to infer beyond the FOV of our
sensor, but many other methods alter the planning and control
pipelines to be reactive to changes in the environment. For
instance [18], [19] focus on quickly planning trajectories
in order to react to new obstacles. In [20], the method
simultaneously plans two trajectories, a safe trajectory in
known space and an aggressive trajectory into unknown
space. For instance, [21] and [22] plan trajectories and verify
safety using only point clouds, similar to [19]. As mentioned
earlier, our strategy differs from these methods as they all
focus on shortening the time between sensor reading and
control generation, whereas our method infers beyond the
sensor FOV to amplify the information available.
III. P RELIMINARIES
A. Problem Formulation
Our objective is to enable an unmanned ground vehicle
(UGV) to navigate, at a high speed to a goal position in an
unknown environment using RGBD and tracking cameras
for perception. Our approach makes use of RGBD mapping,
map prediction, and a receding-horizon controller to achieve
this objective.
For map prediction, the goal of our network architecture
is to learn a function that maps an input occupancy map to
an expanded occupancy map that extends beyond the FOV
of the sensor. More formally, we are learning the function
f : Min 7→ Mout
where Min represents the input occupancy map, and
Mout represents the predicted, expanded output occupancy
map. Components of the function f include an encoder
fenc (Min ) 7→ h ∈ H which maps the input occupancy map
to a hidden state and fdec (h) 7→ (Mout ), which is a decoding

5477

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.

Fig. 3: MIT Racecar with Intel Realsense Cameras
function mapping the hidden state to an expanded, predicted
occupancy map.
The controller accepts robot odometry state, xrobot , the
desired goal, Grobot , a dynamics model of the robot,
VDMrobot , and the predicted occupancy map, Mout to
produce the desired robot velocity, v and turn angle, θ.
B. Platform
The platform we use for our evaluation is the MIT Race
car [23] built on the 1/10-scale Traxxas Rally Car platform,
as shown in Fig. 3. This RC car has a reported maximum
speed of 40 m/s and contains Intel Realsense D435 and T265
cameras. The onboard Nvidia® Jetson TX2 computer runs our
perception, mapping and planning software integrated with
the Robot Operating System (ROS) [24]. Our main interface
to the RC car is the variable electronic speed controller
(VESC) interface that provides vehicle state information and
receives commands including desired velocity and turn angle.
IV. A PPROACH
A. Perception
The objective of the perception algorithm is to observe co-
registered RGB and depth data to produce an occupancy map
for planning. As demonstrated in Fig. 2, RTAB-Map [25] is
used to create 2D occupancy maps using the RGB and depth
images from a Realsense D435 and visual inertial odometry
from a Realsense T265.
To improve our mapping performance, we limit the D435’s
depth sensor range to 3 meters, and we apply gradient
filtering on the raw depth images. The depth image gradients
are calculated using a Sobel filter with a 5 × 5 kernel. All
pixels with a gradient magnitude larger than twice the median
are discarded. This removes ”ghost noise” near sharp edges
in the image. We use RTAB-Map to generate an updated map
at approximately 3 Hz while running on the Nvidia® Jetson
TX2 hardware on the car.
1) Neural Network Architecture: We use a U-Net style
neural network architecture [26], [27] to receive the occu-
pancy map provided by RTAB-Map and predict an expanded
occupancy map. The U-Net neural network is a generative
architecture used in several image completion algorithms
including [27], [28], [29]. The U-Net network consists of
skip connections allowing a direct connection between the
layers i and n − i. These skip connections enable the option
to bypass the bottleneck associated with the downsampling
5478
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
layers and significantly increases the accuracy of predicted
occupancy regions [30]. Our implementation of the encoder
network consists of 7 convolution, batch normalization and
ReLU layers where each convolution consists of a 4 × 4
filter and stride length of 2. The number of filters for the
7 layers in the encoder network are: (64, 128, 256, 512,
512, 512, 512). Similarly, the decoder network consists of 7
upsampling layers with the following number of filters: (512,
1024, 1024, 1024, 512, 256, 128).
2) Loss Function: To train our network, we use a class-
balanced cross-entropy loss function as described in [31].
The discrete classes used for labeling each pixel of the
occupancy map include occupied, unoccupied and unknown
spaces. Because there are significantly more pixels associated
with unoccupied and unknown spaces versus obstacles, we
apply class balancing techniques on the cross entropy loss
with additional 5× weight added to the occupied space
loss. This results in predictions where the edges representing
obstacles are far more pronounced as seen in Fig. 4.
3) Post Processing: During our testing on the robotic car,
we frequently observed small noise artifacts being generated
by the neural network, a condition commonly found in
generative neural networks [32]. While seemingly minor and
transient, these artifacts caused significant instability issues
during control as the trajectory planner would often abruptly
change the planned path in response to these artificial ob-
stacles or would fail to find a valid trajectory. To alleviate
this, we apply a traditional morphological closing operation
with a 5 × 5 kernel to suppress the noise generated by
the neural network with results shown in Fig. 5. While this
closing operation may eliminate pole like objects such as
thin chair legs from the predicted map, overall it produced
a more stable prediction. The observed map and the filtered
predictive map are combined to create the planning map; any
unknown space from the observed map is filled in with data
from the predictive map.
4) Training Details: We generate the datasets in an unsu-
pervised manner. As the robot navigates a new environment,
the robot collects data that consists of a submap that corre-
sponds to the current occupancy map based on the sensor’s
horizon as well as the expanded ground truth map after
the environment has been explored. We explored various
sizes of the submap and found a map corresponding to
6 m × 6 m provided by best geometric size of the submap
given the characteristics of the sensor. We used a map
resolution of 0.05 meters per pixel so the input occupancy
map image resolution was 120 × 120 pixels. Further, we
experimented with various predicted region sizes and found
predicting a region of 7.5 m × 7.5 m corresponding to an
image size of 150 × 150 provided the optimal accuracy and
performance characteristics for the controller. Further, due
to limited amounts of real world data available, we perform
data augmentation techniques to apply random rotations to
the occupancy map training data. This allows us to be more
robust to various hallway configurations as shown in Fig. 6.

Fig. 4: Predicted occupancy map generated without class bal-
ancing weight (Left) and with class balancing weight (Right)
where white represents unoccupied space, grey is occupied
and black is unknown. The class balancing weight produces
stronger edges for obstacles in the predicted occupancy map.
Fig. 5: (Left) Generated occupancy map with noise artifacts.
(Right) Predicted occupancy map after morphological close
operation (white is unoccupied space, grey is occupied and
black is unknown).
B. Control Algorithm
The objective of the control algorithm as described in
Fig. 2 is to compute a collision free path to the goal,
generate a series of feasible trajectories to waypoints, and
send control commands to the mobile robot to follow the
computed trajectory. We adapt the controller proposed in our
prior work, [33]. While initially developed for fixed-wing
flight, we believe it is particularly well suited for high-speed
navigation. The receding horizon allows for rapid replanning
while using a dynamically built map, and the trajectory
generation and tracking allows for a high-rate, dynamically
feasible control output.
1) Dynamics Model: We use a simple bicycle acceleration
model to describe the robot’s dynamics. The equations of
motion are as follows:
ẋ = v ∗ cos(θ)
ẏ = v ∗ sin(θ)
v̇ = u0
θ̇ = v ∗ tan(δ)/L
δ̇ = u1
 
The state is written as x = x, y, v, θ, δ where x and y are
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
2D position, v is forward velocity,
 θ is orientation, and δ is
turn angle. The input, u = u0 , u1 , represents acceleration
and turn angle velocity, respectively, and L is the wheel base
length.
2) Control Strategy: Here, we review the receding horizon
controller proposed in [33], which can be decomposed into
three main stages.
In the first stage, a path to goal is generated using a stan-
dard rapidly-exploring random tree (RRT) [34]. The resulting
path is pruned and then smoothed using G2 Continuous
Cubic Bézier Spiral Path Smoothing  (G2CBS)
 [35]. The
curvature along the smoothed path, x(s), y(s) , is calculated
as:
(y 00 (s)x0 (s) − x00 (s)y 0 (s))
κ(s) = 3
(x0 (s)2 + y 0 (s)2 ) 2
This curvature is then mapped to velocity based upon vmax
and vmin , the maximum and minimum velocity.
dx vmax − vmin
v(s) = (s) = vmax − κ(s) ∗
dt 2
The path’s velocity parameterization is used to reparametrize
the path by time.
Z s
1
t= ds
0 v(s)
In our implementation, we modified the RRT to improve
performance in a dynamically built map by initializing the
RRT tree with the raw RRT path to goal from the previous
control iteration. Before initialization, we check the path
for collisions and truncate it if a collision is detected. This
initialization results in faster RRT computation and increased
path consistency between iterations.
In the second stage, a dynamically feasible trajectory from
the current state to a horizon point is generated. The horizon
point is selected as a time horizon selected along the param-
eterized RRT path. We utilize the same direct transcription
feasibility problem as formulated in [33] and refer the reader
to their formulation. This approach discretizes the trajectory
into N knot points using a variable time interval dt. Let
x0 (tk ) be the position of the robot and u0 (tk ), k < N , be
the input at the k th knot point where tk+1 = tk + dt. We
modify this feasibility problem by introducing a cost function
to penalize large dt (therefore encouraging high speeds) as
well as slightly penalizing the input to encourage smoother
trajectories. The objective function is shown below.
N
X −1
J(x0 (tk ), u0 (tk ), dt) = u0 (tk )T Rc u0 (tk ) + dt
k=0
(1) In order to track the dynamically feasible trajectory, time-
varying LQR (TVLQR) is performed in the third stage. The
control signal is generated as:
u(tk , x) = K(tk )(x − x0 (tk )) + u0 (tk ).
where K(tk ) is the optimal gain matrix.
5479
 Fig. 6: Three examples from our training set of occupancy maps and their resulting expanded predictive map along with
the ground truth (white is unoccupied space, light grey is occupied and dark grey unknown). Augmenting our training data
with random rotations, allows the network prediction to be more robust to different environment configurations encountered
by the robot.

avoidance radius of 0.4 m.

For direct transcription, we used N = 10 knot points
and
 a time horizon of  H = 2 seconds. We set δf =
0.1, 0.1, 0.1, 0.25, 100.0 , Rc = diag(0.1, 0.1) and use an
obstacle radius of 0.35 m.
The costs for TVLQR are as follows:
Q = diag(10, 10, 10, 10, 10)
Qf = diag(1, 1, 5, 1, 1)
R = diag(1, 1)

The acceleration and turn angle control bounds were set to

Fig. 7: Visualization of system during hardware experiment.
Known map is enclosed by the red boundary. The brown
path is the smoothed RRT path to goal, and the purple path
is the local optimized direct transcription trajectory.
3) Control Parameters: The control pipeline executes at
a rate of 5 Hz, or a control interval of T = 0.2 seconds. The
control signal is calculated from the odometry and TVLQR
gains at a rate of 50 Hz. The max time and max iterations for
the RRT search were set to 0.05 seconds and 20000 iterations.
The maximum velocity for RRT path parameterization was
set to the maximum allowed velocity specified in each
hardware trial. For RRT sampling, we use an an obstacle
[-2.5, 2.5] m/s and [-1.5, 1.5] rad/s respectively. The velocity
minimum bound was set to 0.5 m/s and turn angle state
bounds were set to [-0.3, 0.3] rad.
V. E XPERIMENTAL E VALUATION
We conduct preliminary hardware experiments to validate
our approach using a robotic car based on MIT’s open-
source race car [23] as described in III-B. Our map prediction
network was trained on indoor scenes consisting primarily of
straight and turning corridors. Similar to [36], we conduct
both zero-shot and continual learning scenarios. In our zero-
shot experiments, we evaluate the performance on new
environments (Fig. 8) not seen by the robot. In the continual
learning evaluation, we allow the robot to collect data in a
semi-supervised manner from the new environment, fine tune
the network offline and reevaluate performance. We assessed
the maximum speed allowed by the robot and the number of
successful trials, which we defined as reaching the target goal
without collision. In these experiments, unobserved space is
interpreted as unoccupied by the controller. A visualization

5480

Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
 Fig. 8: This sequence of images represents a trajectory taken by the robot during our experimental evaluation.
Algorithm Max Speed Success Rate
Without Map Prediction 3 m/s 5/5
Without Map Prediction 4 m/s 1/5
With Map Prediction (Zero-shot) 4 m/s 3/5
With Map Prediction (Fine Tuned) 4 m/s 4/5
TABLE I: This table captures the results of our preliminary
hardware experiments on our modified MIT race car.
Fig. 9: Example trajectories of car with max velocity of 4
m/s with and without map prediction.
of our system during the hardware experiment is shown in
Fig. 7. The results are summarized in Table I. Without map
prediction, the robot’s maximum speed, vmax , was 3 m/s
without collision. When evaluating without map prediction
with the maximum speed of 4 m/s, the robot was only
able to successfully reach the goal 1/5 attempts. With map
prediction, we were able to achieve success 3/5 trials. After
allowing the network to fine tune on the new environment,
we were able to increase the ratio to 4/5 successful trials
showing the ability to continually learn as the robot explores
new environments. For comparison, the trajectories with and
without prediction for one of the trials where vmax = 4 m/s
is shown in Fig. 9. Without map prediction, the robot, limited
by the sensor’s field of view, optimistically plans a waypoint
in unknown space. At the highest tested velocities, it is not
able to react in time once the map has been updated to reflect
the true occupied space. In contrast, with map prediction, we
are able to plan with longer horizons resulting in smoother
trajectories and more time to react to obstacles which allows
the robot to reach the desired goal.
5481
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
VI. D ISCUSSION AND C ONCLUSION
In this paper, we describe an approach that enables
high speed navigation based on predicted occupancy maps.
Specifically, we present a generative neural network archi-
tecture approach to collect self-supervised training data and
training methodology that allows the robot to predict occu-
pied spaces beyond the line of sight of the camera. Further,
we present several engineering solutions in the perception
algorithm that were required for prediction to work on real
data on the physical hardware including data augmentation,
using a class-balanced loss function and traditional computer
vision methods for noise suppression.
In addition, we also present a real-time controller that
leverages the predicted occupancy map as part of the plan-
ning algorithm. At a max velocity of 3 m/s, 3 Hz mapping
rate, 3 m sensor range, and 1 sec time horizon, planned
trajectories will almost always be within the known region of
the map. It isn’t surprising, thus, that the system is successful
without map prediction with these parameters. When max
velocity is increased to 4 m/s, planned trajectories will often
be within unknown space of the map (outside of the sensor
range). This can cause trajectories to plan through unseen
walls, causing failure. The predicted occupancy map is able
to help address this limitation by providing a longer horizon
for planning which accounts for the improved performance
at 4 m/s in our preliminary hardware evaluation.
While the results are promising, there are many oppor-
tunities for future work. One area is to extend our prelim-
inary hardware experiments to more thoroughly train and
test in various indoor and outdoor environments to assess
the impact of map prediction for high-speed navigation.
No prediction algorithm will be perfect; another area for
future exploration is to incorporate the uncertainty associated
with the occupancy map’s prediction into our controller and
trajectory planner. Possibilities include adjusting the robot’s
velocity based on the the confidence of the prediction, or
rewarding risk-adverse trajectories. Another area to explore is
to improve continual learning techniques as the robot enters
new environments. It is unrealistic to expect the training
data to capture the full distribution of the environments that
the robot will expect to see. We believe further research is
needed to improve the data efficiency of continual learning
techniques so that the robot can improve performance in real
time as it explores new environments.
In spite of these limitations, we have shown the promise of
predictive capabilities that improve navigation performance
of mobile robots and are continually developing techniques
to further extend these capabilities as described above.
 R EFERENCES
[1] H. Voicu and N. Schmajuk, “Exploration, navigation and cognitive
mapping,” Adaptive Behavior, vol. 8, no. 3-4, pp. 207–223, 2000.
[Online]. Available: https://doi.org/10.1177/105971230000800301
[2] R. L. Buckner, “The role of the hippocampus in prediction and
imagination,” Annual review of psychology, vol. 61, pp. 27–48, 2010.
[3] P. Schwartenbeck, J. Passecker, T. Hauser, T. H. B. FitzGerald,
M. Kronbichler, and K. J. Friston, “Computational mechanisms of
curiosity and goal-directed exploration,” bioRxiv, 2018. [Online].
Available: https://www.biorxiv.org/content/early/2018/09/07/411272
[4] S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser,
“Semantic scene completion from a single depth image,” in Pro-
ceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2017, pp. 1746–1754.
[5] H. J. Chang, C. G. Lee, Y.-H. Lu, and Y. C. Hu, “P-slam: Simultaneous
localization and mapping with environmental-structure prediction,”
IEEE Transactions on Robotics, vol. 23, no. 2, pp. 281–293, 2007.
[6] R. Shrestha, F.-P. Tian, W. Feng, P. Tan, and R. Vaughan, “Learned
map prediction for enhanced mobile robot exploration,” in 2019
International Conference on Robotics and Automation (ICRA). IEEE,
2019, pp. 1197–1204.
[7] K. Katyal, K. Popek, C. Paxton, P. Burlina, and G. D. Hager,
“Uncertainty-aware occupancy map prediction using generative net-
works for robot navigation,” in Proceedings - IEEE International
Conference on Robotics and Automation, vol. 2019-May. Institute of
Electrical and Electronics Engineers Inc., may 2019, pp. 5453–5459.
[8] K. D. Katyal, K. M. Popek, C. Paxton, J. L. Moore, K. C.
Wolfe, P. Burlina, and G. D. Hager, “Occupancy map prediction
using generative and fully convolutional networks for vehicle
navigation,” CoRR, vol. abs/1803.02007, 2018. [Online]. Available:
http://arxiv.org/abs/1803.02007
[9] S. K. Ramakrishnan, Z. Al-Halah, and K. Grauman, “Occupancy
anticipation for efficient exploration and navigation,” 2020.
[10] D. P. Ström, F. Nenci, and C. Stachniss, “Predictive exploration consid-
ering previously mapped environments,” in 2015 IEEE international
conference on robotics and automation (ICRA). IEEE, 2015, pp.
2761–2766.
[11] G. J. Stein, C. Bradley, and N. Roy, “Learning over subgoals for effi-
cient navigation of structured, unknown environments,” in Conference
on Robot Learning, 2018, pp. 213–222.
[12] C. Richter, J. Ware, and N. Roy, “High-speed autonomous navigation
of unknown environments using learned probabilities of collision,”
in 2014 IEEE International Conference on Robotics and Automation
(ICRA). IEEE, 2014, pp. 6114–6121.
[13] O. Asraf and V. Indelman, “Experience-based prediction of unknown
environments for enhanced belief space planning,” 10 2020.
[14] A. Elhafsi, B. Ivanovic, L. Janson, and M. Pavone, “Map-Predictive
Motion Planning in Unknown Environments,” in Proceedings - IEEE
International Conference on Robotics and Automation. Institute of
Electrical and Electronics Engineers Inc., may 2020, pp. 8552–8558.
[15] G. Kahn, A. Villaflor, V. Pong, P. Abbeel, and S. Levine,
“Uncertainty-aware reinforcement learning for collision avoidance,”
CoRR, vol. abs/1702.01182, 2017. [Online]. Available: http://arxiv.
org/abs/1702.01182
[16] P. Karkus, D. Hsu, and W. S. Lee, “Qmdp-net: Deep learning for plan-
ning under partial observability,” in Advances in Neural Information
Processing Systems, 2017, pp. 4694–4704.
[17] H. Li, Q. Zhang, and D. Zhao, “Deep reinforcement learning-based
automatic exploration for navigation in unknown environment,” IEEE
transactions on neural networks and learning systems, 2019.
[18] S. Kousik, S. Vaskov, F. Bu, M. Johnson-Roberson, and R. Vasude-
van, “Bridging the gap between safety and real-time performance in
receding-horizon trajectory design for mobile robots,” arXiv preprint
arXiv:1809.06746, 2018.
[19] B. T. Lopez and J. P. How, “Aggressive 3-d collision avoidance for
high-speed navigation.” in ICRA, 2017, pp. 5759–5765.
[20] J. Tordesillas, B. T. Lopez, M. Everett, and J. P. How, “Faster: Fast and
safe trajectory planner for flights in unknown environments,” 2020.
[21] F. Gao, W. Wu, W. Gao, and S. Shen, “Flying on point clouds: Online
trajectory generation and autonomous navigation for quadrotors in
cluttered environments,” Journal of Field Robotics, vol. 36, no. 4,
pp. 710–733, 2019.
[22] P. Florence, J. Carter, and R. Tedrake, “Integrated Perception and
Control at High Speed: Evaluating Collision Avoidance Maneuvers
5482
Authorized licensed use limited to: BEIJING INSTITUTE OF TECHNOLOGY. Downloaded on August 31,2023 at 12:40:34 UTC from IEEE Xplore. Restrictions apply.
Without Maps.” Springer, Cham, 2020, pp. 304–319. [Online].
Available: https://doi.org/10.1007/978-3-030-43089-4{ }20
[23] B. Cain, “High speed autonomous vehicles using the mit racecar
platform,” 2017.
[24] Stanford Artificial Intelligence Laboratory et al., “Robotic operating
system.” [Online]. Available: https://www.ros.org
[25] M. Labbé and F. Michaud, “Rtab-map as an open-source lidar and
visual simultaneous localization and mapping library for large-scale
and long-term online operation,” Journal of Field Robotics, vol. 36,
no. 2, pp. 416–446, 2019.
[26] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional
networks for biomedical image segmentation,” CoRR, vol.
abs/1505.04597, 2015. [Online]. Available: http://arxiv.org/abs/1505.
04597
[27] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image
translation with conditional adversarial networks,” CVPR, 2017.
[28] C. Xie, S. Liu, C. Li, M. Cheng, W. Zuo, X. Liu, S. Wen, and E. Ding,
“Image inpainting with learnable bidirectional attention maps,” in 2019
IEEE/CVF International Conference on Computer Vision, ICCV 2019,
Seoul, Korea (South), October 27 - November 2, 2019. IEEE, 2019,
pp. 8857–8866.
[29] Z. Yan, X. Li, M. Li, W. Zuo, and S. Shan, “Shift-net: Image inpainting
via deep feature rearrangement,” in Computer Vision - ECCV 2018
- 15th European Conference, Munich, Germany, September 8-14,
2018, Proceedings, Part XIV, ser. Lecture Notes in Computer Science,
V. Ferrari, M. Hebert, C. Sminchisescu, and Y. Weiss, Eds., vol. 11218.
Springer, 2018, pp. 3–19.
[30] K. D. Katyal, K. M. Popek, C. Paxton, J. L. Moore, K. C.
Wolfe, P. Burlina, and G. D. Hager, “Occupancy map prediction
using generative and fully convolutional networks for vehicle
navigation,” CoRR, vol. abs/1803.02007, 2018. [Online]. Available:
http://arxiv.org/abs/1803.02007
[31] Y. Cui, M. Jia, T.-Y. Lin, Y. Song, and S. Belongie, “Class-balanced
loss based on effective number of samples,” in CVPR, 2019.
[32] T. Kaneko and T. Harada, “Noise robust generative adversarial net-
works,” in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2020.
[33] M. Basescu and J. Moore, “Direct nmpc for post-stall motion planning
with fixed-wing uavs,” arXiv preprint arXiv:2001.11478, 2020.
[34] S. M. LaValle, “Rapidly-exploring random trees: A new tool for path
planning,” 1998.
[35] K. Yang and S. Sukkarieh, “An analytical continuous-curvature path-
smoothing algorithm,” IEEE Transactions on Robotics, vol. 26, no. 3,
pp. 561–568, 2010.
[36] G. Kahn, P. Abbeel, and S. Levine, “BADGR: an autonomous
self-supervised learning-based navigation system,” CoRR, vol.
abs/2002.05700, 2020. [Online]. Available: https://arxiv.org/abs/2002.
05700